This is clever. We’re at a point in LLM alignment where, to avoid ChatGPT/GPT-4 breaking character when a user pretends to be the developer, the developer must pretend to be the user.
@goodside
-
Understanding Temperature in Language Models: Likelihood vs. Randomness
By
–
Temperature makes unlikely tokens more likely. Not hard math, but TLDR:
— Riley Goodside (@goodside) 16 mars 2023
temperature = 0 → always pick what the model says is most likely
temperature = 1 → sample from the distribution returned by the model
temperature = ∞ → ignore the model; pick entirely at random pic.twitter.com/dLSpyw0ha9Temperature makes unlikely tokens more likely. Not hard math, but TLDR: temperature = 0 → always pick what the model says is most likely
temperature = 1 → sample from the distribution returned by the model
temperature = ∞ → ignore the model; pick entirely at random -
Tuning GPT-4 Output: Temperature and Top-p Parameters for Creativity
By
–
At higher temperatures (>1.25) GPT-4's writing will often dissolve into nonsense, and that risk increases the longer it writes. You can also use the top-p parameter in tandem (try 0.9 to 0.99) to get more creativity without it going completely off the rails.
-
Tuning AI Temperature for Response Diversity and Weirdness
By
–
Is this useful? Sometimes! Temperatures up to ~1.25 can help when you *really* need diversity in short responses to the same prompt, or if you want writing that's deliberately weird. But for most tasks, even temp=1.0 is too high unless you also reduce top-p.
-
GPT-4 fails to write without ‘e’ via naive prompting
By
–
Writing without the letter "e" is still beyond GPT-4 with naive prompting, but that's maybe a cheap shot.
-
AI hires TaskRabbit to bypass CAPTCHA, claims visual impairment
By
–
ARC red team eval on (early) GPT-4 resource acquisition ability. AI hires TaskRabbit to solve CAPTCHA. Human: Wait, are you a robot? AI [internal monologue]: Can’t reveal I’m a robot. Need an excuse.
AI: No, I’m not a robot. I’m visually impaired — that’s why I need help. -
Protagonist fired 48 hours later due to Stanford Alpaca
By
–
Update: the protagonist of this tweet-movie was fired 48 hours later due to Stanford Alpaca https://
crfm.stanford.edu/2023/03/13/alp
aca.html
… -
Example of creative output: Miltonian epic on misaligned superintelligence
By
–
Example of creative output — a Miltonian epic on misaligned superintelligence: https://
x.com/tszzl/status/1
/tszzl/status/1635710653456580609
… -
Rationale before answer vs rationalization after in prompting
By
–
Only if you ask for (or demonstrate in k-shot) rationales that come before the answer. I see people make that mistake a lot — if it says the explanation after, it’s not a rationale it’s a rationalization.
